rl

New

TorchRL is an open-source, PyTorch-native library that provides primitives and higher-level tools to build, train and evaluate reinforcement learning agents and language-model fine-tuning pipelines. It is designed for developers and researchers to implement RL algorithms, manage environments and data, run distributed collectors, and perform modular experiment workflows. The project centers on a Python-first, modular design with a unified data structure called TensorDict that simplifies batched rollouts, replay buffers and training loops. TorchRL also includes an LLM API for supervised fine-tuning and RLHF workflows with conversation management and tool integration. The repository bundles example implementations, SOTA recipes, and utilities for both online and offline RL to accelerate reproducible research and production experimentation.

Stars

3000

App URL

https://github.com/pytorch/rl

Github Repository

https://github.com/pytorch/rl/blob/main/README.md

Features

TorchRL exposes a set of coherent, reusable components: TensorDict for portable tensor-based data handling, a common environment API supporting batched and parallel environments, device-executed environment transforms, synchronous and asynchronous distributed data collectors, and efficient replay buffers including memory-mapped storage and offline dataset wrappers. It provides modular model and exploration wrappers, many loss modules and vectorized advantage computations, a trainer class with hooks, and recipes for common architectures. The LLM API adds unified wrappers for HF and vLLM backends, conversation history management, tool execution transforms, and specialized objectives such as GRPO and SFT. The project is documented, tested, and designed to integrate with the wider PyTorch ecosystem.

Use Cases

TorchRL helps researchers and engineers reduce boilerplate and increase reuse by standardizing data flow and training patterns across environments and algorithms. TensorDict and the environment/transform APIs make it easier to swap sensors, preprocessors and policies without rewriting code. Built-in collectors and multi-process infrastructure accelerate data collection for large-scale experiments. Replay buffers and dataset wrappers support offline RL workflows. Included loss modules, trainers and example SOTA implementations let users prototype and benchmark algorithms quickly. The LLM API supports language-model fine-tuning and RLHF with tool use, enabling experimentation with reward models and specialized objectives. Extensive examples, tutorials and recipes shorten the ramp-up time for new projects and deployments.

rl

Basic Information

Links

Categorization

App Details

Categories

Similar Listings

virtual lab

mnemo

EdgeChains

RAG Agents Accelerator

LLM Zero to Hundred

Featured Listings

Terry Bison Ranch

The Singapore Flyer

Tags

More Filters

rl